Power distribution network fault type identification method based on big data

Abstract

The invention discloses a power distribution network fault type identification method based on big data, and relates to the technical field of power distribution network fault type identification. Multi-source data time axes are aligned on the premise of keeping original sampling precision through clock drift correction and a self-adaptive interpolation algorithm; the problem of misjudgment faults or missed judgment faults caused by asynchronous voltage / current signal time is thoroughly eliminated; low confidence is marked for communication delay data, participation in calculation is limited, and noise interference feature extraction is avoided; on the basis of a fault type association mechanism of a misjudgment sample, a network layer needing to be adjusted is accurately positioned, and resource overhead of full-model retraining is avoided; the scarce fault data is expanded by combining an adversarial sample generation technology, and the recognition robustness of complex scenes such as high-impedance grounding is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault type identification, and particularly to a method for identifying distribution network fault types based on big data. Background Art

[0002] Currently, the identification of distribution network fault types mainly relies on the collaborative analysis of multi-source data such as voltage, current, and communication status. However, there are clock deviations and sampling accuracy differences among different data acquisition devices. For example, the current sampling rate is 1 kHz and the voltage sampling rate is 100 Hz. Direct fusion will cause misalignment of key features.

[0003] Common solutions such as the spatio-temporal graph convolutional network ST-GCN are used to construct a device relationship graph. However, most methods force a unified sampling rate, resulting in the loss of high-frequency details. Moreover, with the continuous access of new elements such as distributed power sources and flexible loads in the distribution network, the distribution of fault features drifts accordingly. For example, after the access of a photovoltaic inverter, the amplitude of the single-phase grounding fault current decreases by 40% - 60%, and the original model misjudges it as a normal state due to outdated training data. Some adaptive solutions, such as incremental learning, can be locally updated, but they rely on retraining by the central server, which takes several hours and cannot meet the requirement of fault handling within minutes.

[0004] There are also some solutions that use edge computing and ST-GCN to retain the spatio-temporal relationship of data locally. However, the graph structure is static and fixed, and nodes cannot be dynamically added or removed, such as the temporary access of a mobile energy storage vehicle. The feedback mechanism mostly uses regular full-scale updates, resulting in large computational overhead and high response latency. It can be seen that the current technology fails to form a closed loop for data alignment - feature extraction - model evolution, and there is an urgent need for a lightweight and adaptive edge intelligence solution. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a method for identifying distribution network fault types based on big data to solve the problems of feature misalignment caused by data asynchrony, model failure due to environmental changes, and difficulty in adapting static solutions to dynamic power grids.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: An embodiment of the present invention provides a method for identifying distribution network fault types based on big data, which includes: Step S1, constructing a dynamic spatio-temporal graph: Real-time collect voltage, current, and communication status data of multiple monitoring points, and construct a spatio-temporal graph with the monitoring point devices as nodes and the physical connections or logical associations between devices as edges; Among them, the original sampling timestamps and precisions of the data of each node are retained, and the time series of different node data are aligned through a time offset compensation mechanism; Step S2, Spatiotemporal Feature Extraction: Input the spatiotemporal graph into the Spatiotemporal Graph Convolutional Network (ST-GCN). Extract the feature of node state change through the convolutional layer in the time dimension, and extract the device association feature through the graph convolutional layer in the space dimension, then output the fused feature vector; Step S3, Fault Classification and Identification: Classify the types of short circuit, grounding, and open circuit faults according to the feature vector, and output the confidence score; Step S4, Feedback Trigger Update: Continuously compare the recognition result with the repair work order record in real time. When the continuous misjudgment reaches the preset number of times or an undefined fault type appears, trigger the update instruction; Step S5, Lightweight Retraining: According to the update instruction, only select the local network layers strongly related to the misjudged samples, and fine-tune the weights with incremental data on the edge nodes.

[0008] As a preferred solution of the method for identifying fault types in a distribution network based on big data according to the present invention, wherein: the time offset compensation mechanism in Step S1 includes: Based on the distribution network synchronous clock source, correct the time drift of the timestamps of each node's data; For data with different sampling precisions, generate a continuous sequence under the unified time axis through the interpolation algorithm; Mark the node data with communication delay exceeding the threshold as low confidence, and restrict its participation in graph convolution calculation.

[0009] As a preferred solution of the method for identifying fault types in a distribution network based on big data according to the present invention, wherein: in Step S1, after clock drift correction and communication delay marking, for the original sampling data of each node Perform the following interpolation processing to construct a unified equidistant time series, including: Detect data: Collect the node The th original moment And the measured value , Calculate the unified time step, the formula is: , wherein, Represents the unified time step, with the unit of second, Represents the th sampling moment of the node , with the unit of second; Construct the aligned time axis , wherein, Represents the th time point after alignment, with the unit of second, Represents the alignment start moment, Represents the sequence length minus one, dimensionless; Calculate the linear interpolation, and the formula is: , where represents the interpolation result of node at time , represents the interpolation weight factor, ; In the formula, , where the numerator is the interpolation time offset, in seconds, and the denominator is the original two - sampling interval; Calculate the interpolation residual within the sliding window: , represents the absolute residual between the th node at the difference time and the reference difference result; if , then adjust or switch to a higher - order interpolation method, where is the higher - order interpolation reference result, is the residual threshold.

[0010] As a preferred solution of the method for identifying distribution network fault types based on big data according to the present invention, wherein: the operations of the spatial - dimensional graph convolutional layer in step S2 include: Dynamically calculate the adjacency matrix weight coefficient based on the electrical distance of the device; For devices with physical connection interruption but logical association, retain the virtual connection edges.

[0011] As a preferred solution of the method for identifying distribution network fault types based on big data according to the present invention, wherein: the selection rules of the local network layer in step S5 include: For misjudgment of short - circuit faults, update the weights of the transient current feature extraction layer in ST - GCN; For misjudgment of grounding faults, update the weights of the zero - sequence voltage feature extraction layer; Determine the network layer number to be updated through the correlation analysis function.

[0012] As a preferred solution of the method for identifying distribution network fault types based on big data according to the present invention, wherein: in step S5, based on the correlation analysis to determine the network layer to be updated, before local re - training, accurately select the network layer number that has the greatest impact on the fault type discrimination from the misjudgment samples, and the steps include: Define the misjudgment sample set , where is the network for the The predicted type of a sample is its true fault type; For each misjudged sample record the output feature map tensor of the th layer of the network where, is the number of channels, is the time series length; Take the norm of the tensor in the two dimensions of channels and time series to obtain the response amplitude of the th layer for the sample : , where, represents the square root norm of the sum of element squares, represents the activation value of the th layer channel and time for the sample ; Define the misjudgment intensity , where, is the confidence score of the network for the sample , obtained from the forward propagation Softmax output, is the misjudgment indicator; Calculate the correlation between the response amplitude of the th layer and the misjudgment intensity. The formula is: , where, is the mean of the sample responses, is the mean of the misjudgments, measures the linear correlation degree between the th layer and the misjudgment; is the adaptive layer selection, introducing the statistic of the global correlation distribution, expressed as: , where, is the total number of candidate update layers, And define the threshold , where, is the threshold adjustment coefficient; Finally, select the layer numbers whose absolute value of the correlation is not less than the threshold: , where, is the set of layer numbers of the network layers whose weights need to be fine-tuned; If, then according to Recalculation in decreasing mode until finally, only perform incremental fine-tuning on the inner layer.

[0013] As a preferred solution of the method for identifying distribution network fault types based on big data according to the present invention, wherein: the incremental data fine-tuning in step S5 includes: Only collect the data of the monitoring points that trigger misjudgment, and the time range covers before and after the fault occurs; Introduce the adversarial sample generation technology to expand the data of scarce fault scenarios such as high-impedance grounding.

[0014] As a preferred solution of the method for identifying distribution network fault types based on big data according to the present invention, wherein: the dynamic maintenance of the spatio-temporal graph in step S1 further includes: When adding a distributed power generation node, automatically inherit the topological relationship of the associated feeder; For offline device nodes, retain historical data but block the input of real-time data streams.

[0015] As a preferred solution of the method for identifying distribution network fault types based on big data according to the present invention, wherein: the fault classification and identification in step S3 includes: When the confidence score is lower than the threshold, jointly make a collaborative decision with the feature vectors of the adjacent 3 monitoring points; For undefined fault types, extract their spatio-temporal feature vectors and store them in the template library and alarm.

[0016] As a preferred solution of the method for identifying distribution network fault types based on big data according to the present invention, wherein: after triggering the update instruction in step S4: Generate a model update container at the edge node to isolate the retraining process from the online recognition service; After the update is completed, perform cross-validation, and after passing the validation, hot-switch to the online recognition module.

[0017] The beneficial effects of the present invention are as follows: In the present invention, through the clock drift correction and adaptive interpolation algorithm, the time axes of multi-source data are aligned on the premise of retaining the original sampling accuracy, completely eliminating the problems of misjudged faults or missed judged faults caused by the asynchronous time of voltage / current signals; mark the communication delay data with low confidence and limit its participation in calculations to avoid noise interference in feature extraction.

[0018] Based on the fault type association mechanism of misjudged samples (such as updating the current layer in case of short-circuit misjudgment), accurately locate the network layer to be adjusted, avoiding the resource overhead of full model retraining; combine the adversarial sample generation technology to expand scarce fault data and improve the recognition robustness for complex scenarios such as high-impedance grounding.

[0019] Lightweight retraining only requires the resources of edge nodes. The model update process is isolated from the online recognition service to ensure uninterrupted fault recognition. The dynamic spatio-temporal graph structure automatically adapts to the topological changes of the distribution network (such as the addition of photovoltaic devices), eliminating the need for manual model reconstruction.

[0020] The confidence score links multi-monitoring points for collaborative decision-making, reducing the frequency of manual review. The undefined fault features are automatically stored in the database for alarm, accelerating the disposal loop of new fault types.

[0021] In the present invention, data alignment - feature extraction - model evolution form a closed loop, achieving the adaptability, accuracy, and efficiency of distribution network fault identification and discrimination with edge intelligence. Description of the Drawings

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a schematic flowchart of the method for identifying distribution network fault types based on big data in Embodiment 1. Detailed Embodiments

[0024] To make the above objects, features, and advantages of the present invention more obvious and understandable, the detailed embodiments of the present invention will be described in detail below with reference to the drawings of the specification.

[0025] Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0026] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that excludes other embodiments.

[0027] Embodiment 1, referring to Figure 1 , this embodiment provides a method for identifying distribution network fault types based on big data, including the following steps: Step S1, constructing a dynamic spatio-temporal graph: Real-time collect the voltage, current, and communication status data of multiple monitoring points, and construct a spatio-temporal graph with the monitoring point devices as nodes and the physical connections or logical associations between devices as edges; The original sampling timestamps and precisions of each node's data are retained, and the time series of different nodes' data are aligned through a time offset compensation mechanism; The time offset compensation mechanism in step S1 includes: Taking the distribution network synchronous clock source as a reference, drift correction is performed on the timestamps of each node's data; For data with different sampling precisions, a continuous sequence under a unified time axis is generated through an interpolation algorithm; Node data with a communication delay exceeding the threshold is marked as low confidence, and its participation in graph convolution calculation is restricted; In step S1, after clock drift correction and communication delay marking, for the original sampling data of each node the following interpolation processing is performed to construct a unified equidistant time series, including: Detecting data: Collecting the node at the th original moment and the measured value , Calculating the unified time step, the formula is: , where represents the unified time step, with the unit of seconds, represents the th sampling moment of the node, with the unit of seconds; Constructing the aligned time axis , , where represents the th time point after alignment, with the unit of seconds, represents the alignment start moment, represents the sequence length minus one, dimensionless; Calculating the linear interpolation, the formula is: , where represents the interpolation result of the node at the moment , represents the interpolation weight factor, ; In the formula, , where the numerator is the interpolation moment offset, with the unit of seconds, and the denominator is the original sampling interval between two times; Calculating the interpolation residual within the sliding window: , Indicates The node is The absolute residual between the difference moment and the reference difference result; if , then adjust Or switch to a higher-order interpolation method, where is the high-order interpolation reference result, is the residual threshold; Specifically, the interpolation process generates a global unified time axis through the minimum step method while maintaining the original sampling time and data integrity, avoiding the loss of high-frequency information. The linear interpolation calculation is concise and easy to execute in real time on edge devices. The weight factor accurately reflects the time interval, which helps to smoothly transition to the downstream spatiotemporal graph convolutional network. The sliding window residual judgment and dynamic step adjustment mechanism improve the compatibility with data of various sampling precisions and avoid oversampling or undersampling problems caused by fixed step sizes. The dynamic maintenance of the spatiotemporal graph in step S1 also includes: When adding a distributed power supply node, the topological relationship of the associated feeder is automatically inherited; For offline device nodes, historical data is retained but real-time data stream input is blocked; Step S2, spatiotemporal feature extraction: input the spatiotemporal graph into the spatiotemporal graph convolutional network ST-GCN, extract the node state change features through the time dimension convolution layer, extract the device association features through the space dimension convolution layer, and output the fusion feature vector; The operations of the spatial dimension graph convolution layer in step S2 include: Dynamically calculate the adjacency matrix weight coefficient based on the electrical distance of the equipment; For devices that are physically disconnected but logically associated (such as the same feeder partition), the virtual connection edge is retained; Step S3, fault classification and identification: classify the short circuit, grounding, and disconnection fault types according to the feature vector, and output the confidence score; The fault classification identification in step S3 includes: When the confidence score is lower than the threshold, the feature vectors of the three adjacent monitoring points are combined for collaborative decision-making; For undefined fault types, extract their spatiotemporal feature vectors, store them in the template library and generate an alarm; Step S4, feedback trigger update: real-time comparison of recognition results and emergency repair work order records, when continuous misjudgment reaches a preset number of times or an undefined fault type occurs, trigger an update instruction; After the update instruction is triggered in step S4: Generate a model update container on the edge node to isolate the retraining process from the online recognition service; After the update is completed, cross-verification is performed, and after verification is passed, hot switching to the online recognition module; Step S5, lightweight retraining: According to the update instruction, only select the local network layers strongly related to the misjudged samples, and fine-tune the weights using incremental data at the edge nodes; The local network layer selection rules in step S5 include: For misjudgment of short-circuit faults, update the weights of the transient current feature extraction layer in ST-GCN; For misjudgment of grounding faults, update the weights of the zero-sequence voltage feature extraction layer; Determine the network layer numbers to be updated through the correlation analysis function; In step S5, based on the correlation analysis to determine the network layers to be updated, before local retraining, accurately select the network layer numbers that have the greatest impact on the fault type discrimination from the misjudged samples. The steps include: Define the misjudged sample set , where is the predicted type of the network for the th sample, is its true fault type; For each misjudged sample , record the output feature map tensor of the th layer of the network , where is the number of channels, is the time series length; Take the norm of the tensor in both the channel and time series dimensions to obtain the response amplitude of the th layer for the sample : , where represents the square root of the sum of squares of elements norm, represents the activation value of the th layer channel , time for the sample ; Define the misjudgment intensity , where is the confidence score of the network for the sample , obtained from the forward propagation Softmax output, is the misjudgment indicator; Calculate the correlation between the response amplitude of the th layer and the misjudgment intensity. The formula is: , where , sample response mean, , misjudgment mean, , measure the The linear correlation degree with misjudgment; For adaptive layer selection, a statistic representing the global correlation distribution is introduced, denoted as: , where, is the total number of candidate update layers, and a threshold is defined, where is the threshold adjustment coefficient; Finally, select the layer numbers whose absolute correlation value is not less than the threshold: , where is the set of network layer numbers whose weights need to be fine-tuned; If, then recalculate in a decreasing manner until , and finally only perform incremental fine-tuning on the layers within, reducing the edge computing burden; Specifically, this method utilizes the statistical correlation between the network responses of misjudged samples and the misjudgment intensity to accurately identify the network layers that are most sensitive to fault type discrimination. Through norm dimensionality reduction and Pearson correlation coefficient calculation, it not only retains the overall energy information of the feature mapping but also reveals the direct impact of the layer on errors. The dynamic correlation threshold combined with the global distribution statistics overcomes the non-robustness of a single fixed threshold, ensuring that effective layers can be selected in different misjudgment scenarios. The mechanism of decreasing the threshold ensures that key layers are not missed in cases of insufficient or scattered samples. The overall algorithm only involves one-dimensional vector operations and simple statistics, with extremely high computational efficiency and is suitable for real-time deployment at resource-constrained edge nodes; The incremental data fine-tuning in step S5 includes: Only collect the data of the monitoring points that trigger misjudgment, and the time range covers before and after the fault occurs; Introduce adversarial sample generation technology to expand data for scarce fault scenarios such as high-impedance grounding.

[0028] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.​

Claims

1. A method for identifying distribution network fault types based on big data, characterized in that, including Step S1, constructing a dynamic spatio-temporal graph: Collecting voltage, current, and communication status data of multiple monitoring points in real time, and constructing a spatio-temporal graph with the monitoring point devices as nodes and the physical connections or logical associations between devices as edges; wherein the original sampling timestamps and precisions of the data of each node are retained, and the time series of the data of different nodes are aligned through a time offset compensation mechanism; Step S2, spatio-temporal feature extraction: Inputting the spatio-temporal graph into the spatio-temporal graph convolutional network ST-GCN, extracting the characteristics of node state changes through the convolutional layer in the time dimension, and extracting the device association characteristics through the graph convolutional layer in the space dimension, and outputting a fused feature vector; Step S3, fault classification and identification: Classifying short-circuit, grounding, and disconnection fault types according to the feature vector, and outputting a confidence score; Step S4, feedback trigger update: Comparing the recognition result with the repair work order record in real time. When the continuous misjudgment reaches the preset number of times or an undefined fault type appears, a trigger update instruction is issued; Step S5, lightweight retraining: According to the update instruction, only select the local network layers strongly related to the misjudged samples, and fine-tune the weights with incremental data at the edge nodes.

2. The method for identifying the fault type of a distribution network based on big data according to claim 1, wherein The time offset compensation mechanism in Step S1 includes: Taking the distribution network synchronous clock source as the reference, correcting the drift of the timestamps of the data of each node; For data with different sampling precisions, generating a continuous sequence under a unified time axis through an interpolation algorithm; Marking the data of the nodes with communication delays exceeding the threshold as low confidence, and restricting their participation in graph convolution calculations.

3. The method for identifying the type of distribution network fault based on big data according to claim 2, characterized in that In step S1, after clock drift correction and communication delay marking, for the original sampling data of each node perform the following interpolation processing to construct a unified equidistant time series, including: Detection data: Acquisition node The th original moment and the measured value , Calculating the unified time step, the formula is: ， Among them, represents the unified time step, with the unit of second, represents the node the th sampling moment, with the unit of second; Build an aligned timeline , Among them, represents the th time point after alignment, with the unit of second, represents the starting moment of alignment, represents the sequence length minus one, dimensionless; Calculating the linear interpolation, the formula is: ， Among them, represents the node at the moment interpolation result, represents the interpolation weight factor, ; In the formula, ， Among them, the numerator is the interpolation time offset, with the unit of second, and the denominator is the original sampling interval between two times; Calculating the interpolation residual within the sliding window: , represents the node's absolute residual between the difference moment and the reference difference result; if , then adjust or switch to a higher-order interpolation method, where is the higher-order interpolation reference result, is the residual threshold.

4. The method for identifying a distribution network fault type based on big data according to claim 1, characterized in that The operations of the graph convolutional layer in the space dimension in Step S2 include: Dynamically calculating the weight coefficients of the adjacency matrix based on the electrical distance of the devices; For devices with physical connections interrupted but logical associations, retaining the virtual connection edges.

5. A method for identifying the fault type of a distribution network based on big data according to claim 1, characterized in that, The selection rules for the local network layers in Step S5 include: For misjudgment of short-circuit faults, updating the weights of the transient current feature extraction layer in ST-GCN; For misjudgment of grounding faults, updating the weights of the zero-sequence voltage feature extraction layer; Determining the numbers of the network layers to be updated through a correlation analysis function.

6. The method for identifying the fault type of a distribution network based on big data according to claim 5, wherein In Step S5, based on the correlation analysis to determine the network layers to be updated. Before local retraining, accurately select the numbers of the network layers that have the greatest impact on the discrimination of fault types from the misjudged samples. The steps include: Define the misjudgment sample set , where is the predicted type of the network for the th sample, and is its true fault type; For each misjudged sample , record the output feature map tensor of the layer of the network. Among them, is the number of channels, is the time series length; Take the tensor and take the norm in the two dimensions of channel and time sequence to obtain layer for the sample to get the response amplitude: ， Among them, represents the square root norm of the sum of element squares, represents the layer channel , at time the activation value of the sample . Define the misjudgment intensity , Among them, is the confidence score of the network for the sample obtained from the forward propagation Softmax output, is the misjudgment indicator; Computing layer The correlation between the response amplitude and the misjudgment intensity is expressed by the formula: ， Among them, , the sample response mean, , the misjudgment mean, , measure layer and the linear correlation degree between misjudgments; For adaptive layer selection, introducing a statistic of the global correlation distribution, expressed as: ， Among them, is the total number of candidate update layers, And define a threshold value , where is the threshold adjustment coefficient; Finally, selecting the layer numbers with the absolute value of the correlation not less than the threshold: ， Among them, is the set of network layer numbers whose weights need to be fine-tuned; If then recalculate according to the decreasing method until and finally perform incremental fine-tuning only on the inner layers.

7. The method for identifying the fault type of a distribution network based on big data according to claim 1, wherein, The incremental data fine-tuning in Step S5 includes: Only collecting the data of the monitoring points that trigger misjudgment, and the time range covers before and after the fault occurs; Introducing the adversarial sample generation technology to expand the data of scarce fault scenarios such as high-impedance grounding.

8. A method for identifying distribution network fault types based on big data according to claim 1, characterized in that, The dynamic maintenance of the spatio-temporal graph in Step S1 also includes: When adding distributed power generation nodes, automatically inheriting the topological relationship of the associated feeders; For offline device nodes, retaining historical data but blocking the input of real-time data streams.

9. The method for identifying the fault type of a distribution network based on big data according to claim 1, wherein, The fault classification and identification in Step S3 include: When the confidence score is lower than the threshold, jointly making a decision with the feature vectors of the adjacent 3 monitoring points; For undefined fault types, extracting their spatio-temporal feature vectors and storing them in the template library and alarming.

10. A method for identifying distribution network fault types based on big data according to claim 1, characterized in that, After issuing the trigger update instruction in Step S4: Generate a model update container at the edge node to isolate the retraining process from the online recognition service; After the update is completed, perform cross-validation. After passing the validation, hot-switch to the online recognition module.

Images